automatic import of ANNC

3 files changed, 194 insertions, 158 deletions

diff --git a/0003-Add-python-bindings-for-kernels.patch b/0003-Add-python-bindings-for-kernels.patch
new file mode 100644
index 0000000..22f70dc
--- /dev/null
+++ b/0003-Add-python-bindings-for-kernels.patch
@@ -0,0 +1,192 @@
+From 97c972045ab8bf8f4f4db39da0728b2c41183eff Mon Sep 17 00:00:00 2001
+From: zhengchenhui <zhengchenhui1@huawei.com>
+Date: Wed, 12 Nov 2025 16:38:50 +0800
+Subject: [PATCH] Add python bindings for kernels.
+
+---
+ ...i_def_KPFusedEmbeddingActionIdGather.pbtxt |  3 ++
+ .../api_def_KPFusedEmbeddingPadding.pbtxt     |  3 ++
+ .../api_def_KPFusedEmbeddingPaddingFast.pbtxt |  3 ++
+ .../api_def/api_def_KPFusedGather.pbtxt       |  3 ++
+ .../api_def_KPFusedSparseDynamicStitch.pbtxt  |  3 ++
+ .../api_def_KPFusedSparseReshape.pbtxt        |  3 ++
+ .../api_def_KPFusedSparseSegmentReduce.pbtxt  |  3 ++
+ ...ef_KPFusedSparseSegmentReduceNonzero.pbtxt |  3 ++
+ .../api_def/api_def_KPFusedSparseSelect.pbtxt |  3 ++
+ annc/tensorflow/tf_annc_optimizer.patch       | 53 +++++++++++++++++--
+ 10 files changed, 76 insertions(+), 4 deletions(-)
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedEmbeddingActionIdGather.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedEmbeddingPadding.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedEmbeddingPaddingFast.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedGather.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedSparseDynamicStitch.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedSparseReshape.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduce.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduceNonzero.pbtxt
+ create mode 100644 annc/tensorflow/api_def/api_def_KPFusedSparseSelect.pbtxt
+
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedEmbeddingActionIdGather.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedEmbeddingActionIdGather.pbtxt
+new file mode 100644
+index 0000000..bb221ad
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedEmbeddingActionIdGather.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedEmbeddingActionIdGather"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedEmbeddingPadding.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedEmbeddingPadding.pbtxt
+new file mode 100644
+index 0000000..5aef378
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedEmbeddingPadding.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedEmbeddingPadding"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedEmbeddingPaddingFast.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedEmbeddingPaddingFast.pbtxt
+new file mode 100644
+index 0000000..17e7ebc
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedEmbeddingPaddingFast.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedEmbeddingPaddingFast"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedGather.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedGather.pbtxt
+new file mode 100644
+index 0000000..63c4eef
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedGather.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedGather"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedSparseDynamicStitch.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedSparseDynamicStitch.pbtxt
+new file mode 100644
+index 0000000..5b01d70
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedSparseDynamicStitch.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedSparseDynamicStitch"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedSparseReshape.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedSparseReshape.pbtxt
+new file mode 100644
+index 0000000..8fe3c2f
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedSparseReshape.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedSparseReshape"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduce.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduce.pbtxt
+new file mode 100644
+index 0000000..b13aa7d
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduce.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedSparseSegmentReduce"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduceNonzero.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduceNonzero.pbtxt
+new file mode 100644
+index 0000000..c0945af
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedSparseSegmentReduceNonzero.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedSparseSegmentReduceNonzero"
++}
+diff --git a/annc/tensorflow/api_def/api_def_KPFusedSparseSelect.pbtxt b/annc/tensorflow/api_def/api_def_KPFusedSparseSelect.pbtxt
+new file mode 100644
+index 0000000..bafc915
+--- /dev/null
++++ b/annc/tensorflow/api_def/api_def_KPFusedSparseSelect.pbtxt
+@@ -0,0 +1,3 @@
++op {
++  graph_op_name: "KPFusedSparseSelect"
++}
+diff --git a/annc/tensorflow/tf_annc_optimizer.patch b/annc/tensorflow/tf_annc_optimizer.patch
+index c92330a..ec0199e 100644
+--- a/annc/tensorflow/tf_annc_optimizer.patch
++++ b/annc/tensorflow/tf_annc_optimizer.patch
+@@ -1,5 +1,5 @@
+ diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD
+-index 538574360ba..b43e0455802 100644
++index 538574360ba..3864064d18c 100644
+ --- a/tensorflow/core/BUILD
+ +++ b/tensorflow/core/BUILD
+ @@ -629,6 +629,7 @@ cc_library(
+@@ -10,6 +10,14 @@ index 538574360ba..b43e0455802 100644
+      ] + if_mkl([
+          "//tensorflow/core/kernels/mkl:mkl_concat_op",
+          "//tensorflow/core/kernels/mkl:mkl_dequantize_op",
++@@ -907,6 +908,7 @@ filegroup(
++         "candidate_sampling_ops_op_lib",
++         "checkpoint_ops_op_lib",
++         "clustering_ops_op_lib",
+++        "embedding_fused_ops_op_lib",
++         "collective_ops_op_lib",
++         "control_flow_ops_op_lib",
++         "count_ops_op_lib",
+ diff --git a/tensorflow/core/grappler/optimizers/BUILD b/tensorflow/core/grappler/optimizers/BUILD
+ index ecd559734ea..97a918ead6d 100644
+ --- a/tensorflow/core/grappler/optimizers/BUILD
+@@ -189,6 +197,45 @@ index 91d80b6c2b5..b00b2a5d027 100644
+      ] + select({
+          # Non-tpu platforms don't need tpu dependency.
+          "//tensorflow:chromiumos": [],
++diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD
++index 9810f8acd2e..3b0b23fec0a 100644
++--- a/tensorflow/python/BUILD
+++++ b/tensorflow/python/BUILD
++@@ -269,6 +269,7 @@ py_strict_library(
++         "//tensorflow/python/ops/signal:util_ops",
++         "//tensorflow/python/ops/signal:window_ops",
++         "//tensorflow/python/ops/structured:structured_ops",
+++        "//tensorflow/python/ops:embedding_fused_ops_gen",
++         "//tensorflow/python/platform:_pywrap_stacktrace_handler",
++         "//tensorflow/python/platform:app",
++         "//tensorflow/python/platform:client_testlib",
++@@ -389,6 +390,7 @@ py_strict_library(
++         "//tensorflow/python/ops:bitwise_ops",
++         "//tensorflow/python/ops:boosted_trees_ops_gen",
++         "//tensorflow/python/ops:clustering_ops_gen",
+++        "//tensorflow/python/ops:embedding_fused_ops_gen",
++         "//tensorflow/python/ops:composite_tensor_ops",
++         "//tensorflow/python/ops:cond_v2",
++         "//tensorflow/python/ops:cudnn_rnn_ops_gen",
++diff --git a/tensorflow/python/ops/BUILD b/tensorflow/python/ops/BUILD
++index b3a3d612a01..185da7da4a8 100644
++--- a/tensorflow/python/ops/BUILD
+++++ b/tensorflow/python/ops/BUILD
++@@ -228,6 +228,14 @@ tf_gen_op_strict_wrapper_private_py(
++     ],
++ )
++
+++tf_gen_op_strict_wrapper_private_py(
+++    name = "embedding_fused_ops_gen",
+++    visibility = ["//tensorflow:internal"],
+++    deps = [
+++        "//tensorflow/core:embedding_fused_ops_op_lib",
+++    ],
+++)
+++
++ tf_gen_op_strict_wrapper_private_py(
++     name = "collective_ops_gen",
++     visibility = ["//tensorflow:internal"],
+ diff --git a/third_party/xla/xla/service/cpu/BUILD b/third_party/xla/xla/service/cpu/BUILD
+ index 6e0ea613435..47c346b4e93 100644
+ --- a/third_party/xla/xla/service/cpu/BUILD
+@@ -216,6 +263,4 @@ index 6e0ea613435..47c346b4e93 100644
+ +
+  cc_library(
+      name = "cpu_compiler_pure",
+-     srcs = ["cpu_compiler.cc"],
+-
+-
++     srcs = ["cpu_compiler.cc"],
+\ No newline at end of file
+-- 
+2.33.0
+
diff --git a/ANNC.spec b/ANNC.spec
index ca85e5a..6b3ca7d 100644
--- a/ANNC.spec
+++ b/ANNC.spec
@@ -28,6 +28,7 @@ Patch0:    x86_64_external_files.patch
 %endif
 Patch1:     0001-fix-pattern-conflicts.patch
 Patch2:     0002-Add-graph-optimizer-and-embedding_fused-kernels.patch
+Patch3:     0003-Add-python-bindings-for-kernels.patch
 
 BuildRequires:  cmake >= 3.9.9
 BuildRequires:  make
@@ -60,6 +61,7 @@ cp %{SOURCE5} ./proxy
 %setup -q -n %{pkg_version}
 %patch1 -p1
 %patch2 -p1
+%patch3 -p1
 
 %build
 export ANNC=%{build_dir}
diff --git a/x86_64_external_files.patch b/x86_64_external_files.patch
deleted file mode 100644
index 7bf4fb3..0000000
--- a/x86_64_external_files.patch
+++ /dev/null
@@ -1,158 +0,0 @@
-diff --git a/external/local_config_cc/BUILD b/external/local_config_cc/BUILD
-index 6c19fae..a0ea3f8 100755
---- a/external/local_config_cc/BUILD
-+++ b/external/local_config_cc/BUILD
-@@ -47,15 +47,15 @@ filegroup(
- cc_toolchain_suite(
-     name = "toolchain",
-     toolchains = {
--        "aarch64|compiler": ":cc-compiler-aarch64",
--        "aarch64": ":cc-compiler-aarch64",
-+        "k8|compiler": ":cc-compiler-k8",
-+        "k8": ":cc-compiler-k8",
-         "armeabi-v7a|compiler": ":cc-compiler-armeabi-v7a",
-         "armeabi-v7a": ":cc-compiler-armeabi-v7a",
-     },
- )
- 
- cc_toolchain(
--    name = "cc-compiler-aarch64",
-+    name = "cc-compiler-k8",
-     toolchain_identifier = "local",
-     toolchain_config = ":local",
-     all_files = ":compiler_deps",
-@@ -72,7 +72,7 @@ cc_toolchain(
- 
- cc_toolchain_config(
-     name = "local",
--    cpu = "aarch64",
-+    cpu = "k8",
-     compiler = "compiler",
-     toolchain_identifier = "local",
-     host_system_name = "local",
-@@ -80,7 +80,7 @@ cc_toolchain_config(
-     target_libc = "local",
-     abi_version = "local",
-     abi_libc_version = "local",
--    cxx_builtin_include_directories = ["/usr/lib/gcc/aarch64-linux-gnu/10.3.1/include",
-+    cxx_builtin_include_directories = ["/usr/lib/gcc/x86_64-linux-gnu/10.3.1/include",
-     "/usr/local/include",
-     "/usr/include",
-     "/root/rpmbuild/BUILD/ANNC-v0.0.2/annc/service/cpu/xla",
-@@ -91,7 +91,7 @@ cc_toolchain_config(
-     "/root/rpmbuild/BUILD/XNNPACK/build/pthreadpool-source/include",
-     "/root/rpmbuild/BUILD/ANNC-v0.0.2/output/e7b069029cc648c50e1b8083cef52b4f/external/local_config_cc",
-     "/usr/include/c++/10.3.1",
--    "/usr/include/c++/10.3.1/aarch64-linux-gnu",
-+    "/usr/include/c++/10.3.1/x86_64-linux-gnu",
-     "/usr/include/c++/10.3.1/backward"],
-     tool_paths = {"ar": "/usr/bin/ar",
-         "ld": "/usr/bin/ld",
-diff --git a/external/local_config_cc/builtin_include_directory_paths b/external/local_config_cc/builtin_include_directory_paths
-index 188e6c9..2c3bcc1 100755
---- a/external/local_config_cc/builtin_include_directory_paths
-+++ b/external/local_config_cc/builtin_include_directory_paths
-@@ -4,7 +4,7 @@ changes to it will be reflected in the action cache key. When some of these
- paths change, Bazel will make sure to rerun the action, even though none of
- declared action inputs or the action commandline changes.
- 
--/usr/lib/gcc/aarch64-linux-gnu/10.3.1/include
-+/usr/lib/gcc/x86_64-linux-gnu/10.3.1/include
- /usr/local/include
- /usr/include
- /root/rpmbuild/BUILD/ANNC-v0.0.2/annc/service/cpu/xla
-@@ -15,5 +15,5 @@ declared action inputs or the action commandline changes.
- /root/rpmbuild/BUILD/XNNPACK/build/pthreadpool-source/include
- /root/rpmbuild/BUILD/ANNC-v0.0.2/output/e7b069029cc648c50e1b8083cef52b4f/external/local_config_cc
- /usr/include/c++/10.3.1
--/usr/include/c++/10.3.1/aarch64-linux-gnu
-+/usr/include/c++/10.3.1/x86_64-linux-gnu
- /usr/include/c++/10.3.1/backward
-diff --git a/external/go_sdk/BUILD.bazel b/x86_64/external/go_sdk/BUILD.bazel
-index 9cf6add..511ddbc 100644
---- a/external/go_sdk/BUILD.bazel
-+++ b/x86_64/external/go_sdk/BUILD.bazel
-@@ -7,8 +7,8 @@ package(default_visibility = ["//visibility:public"])
- filegroup(
-     name = "libs",
-     srcs = glob(
--        ["pkg/linux_arm64/**/*.a"],
--        exclude = ["pkg/linux_arm64/**/cmd/**"],
-+        ["pkg/linux_amd64/**/*.a"],
-+        exclude = ["pkg/linux_amd64/**/cmd/**"],
-     ),
- )
- 
-@@ -34,7 +34,7 @@ go_sdk(
-     name = "go_sdk",
-     srcs = [":srcs"],
-     go = "bin/go",
--    goarch = "arm64",
-+    goarch = "amd64",
-     goos = "linux",
-     headers = [":headers"],
-     libs = [":libs"],
-@@ -60,7 +60,7 @@ package_list(
- 
- declare_toolchains(
-     builder = ":builder",
--    host = "linux_arm64",
-+    host = "linux_amd64",
-     sdk = ":go_sdk",
- )
- 
-diff --git a/external/go_sdk/src/internal/buildcfg/zbootstrap.go b/x86_64/external/go_sdk/src/internal/buildcfg/zbootstrap.go
-index 81bd76b..560f94f 100644
---- a/external/go_sdk/src/internal/buildcfg/zbootstrap.go
-+++ b/x86_64/external/go_sdk/src/internal/buildcfg/zbootstrap.go
-@@ -12,7 +12,7 @@ const defaultGOMIPS64 = `hardfloat`
- const defaultGOPPC64 = `power8`
- const defaultGOEXPERIMENT = ``
- const defaultGO_EXTLINK_ENABLED = ``
--const defaultGO_LDSO = `/lib/ld-linux-aarch64.so.1`
-+const defaultGO_LDSO = `/lib64/ld-linux-x86-64.so.2`
- const version = `go1.18.4`
- const defaultGOOS = runtime.GOOS
- const defaultGOARCH = runtime.GOARCH
-diff --git a/external/local_config_cc_toolchains/BUILD b/x86_64/external/local_config_cc_toolchains/BUILD
-index db5234f..f9c0875 100755
---- a/external/local_config_cc_toolchains/BUILD
-+++ b/x86_64/external/local_config_cc_toolchains/BUILD
-@@ -1,10 +1,10 @@
- load("@local_config_platform//:constraints.bzl", "HOST_CONSTRAINTS")
- 
- toolchain(
--    name = "cc-toolchain-aarch64",
-+    name = "cc-toolchain-k8",
-     exec_compatible_with = HOST_CONSTRAINTS,
-     target_compatible_with = HOST_CONSTRAINTS,
--    toolchain = "@local_config_cc//:cc-compiler-aarch64",
-+    toolchain = "@local_config_cc//:cc-compiler-k8",
-     toolchain_type = "@bazel_tools//tools/cpp:toolchain_type",
- )
- 
-diff --git a/external/local_config_platform/constraints.bzl b/x86_64/external/local_config_platform/constraints.bzl
-index 4b29ad7..6a7d8c3 100644
---- a/external/local_config_platform/constraints.bzl
-+++ b/x86_64/external/local_config_platform/constraints.bzl
-@@ -1,6 +1,6 @@
- # DO NOT EDIT: automatically generated constraints list for local_config_platform
- # Auto-detected host platform constraints.
- HOST_CONSTRAINTS = [
--  '@platforms//cpu:aarch64',
-+  '@platforms//cpu:x86_64',
-   '@platforms//os:linux',
- ]
-diff --git a/external/local_execution_config_platform/BUILD b/x86_64/external/local_execution_config_platform/BUILD
-index b4d7fdf..7dfbc12 100755
---- a/external/local_execution_config_platform/BUILD
-+++ b/x86_64/external/local_execution_config_platform/BUILD
-@@ -18,7 +18,7 @@ platform(
-     name = "platform",
-     visibility = ["//visibility:public"],
-     constraint_values = [
--        "@platforms//cpu:aarch64",
-+        "@platforms//cpu:x86_64",
-         "@platforms//os:linux",
-         ":platform_constraint",
-     ],